Willis Wu.
The project collection

08 / Software · Robotics

Traffic sign recognition

I built a YOLOv11 detection and PyTorch classification pipeline for traffic-sign recognition.

My role

Detection/classification pipeline, dataset preparation, and validation

Context

Computer vision project

Dates

May 2025 – Aug 2025

Inference demonstration

Screen-recorded inference on daytime and nighttime road scenes.

Predicted sign regions, labels and confidence annotations update as each scene progresses.

Training images
50k+
Mapillary Traffic Sign Dataset

01 / The brief

Localize first, classify second.

I used YOLOv11 to detect candidate sign regions and a custom PyTorch CNN to classify them. Separating the stages made dataset preparation and class imbalance explicit parts of the pipeline.

02 / My contribution

Dataset preparation and reproducible runs.

I trained on 50k+ Mapillary Traffic Sign Dataset images, cleaned labels and applied targeted augmentation for class imbalance. I used Docker/CUDA workflows for reproducible training and validation, and tested webcam and vehicle-captured video under day/night conditions.

03 / Evidence & scope

Actual predictions on street imagery.

These inference frames show the pipeline’s predicted boxes, labels and confidence annotations on street scenes.

Inference output

03
Street image with green model-predicted bounding boxes and traffic-sign labels.
Model predictionsStreet-scene inference

Actual inference output. Green boxes and labels show the model’s predictions on street imagery.

View original
City street inference output with multiple predicted sign regions and labels.
Model predictionsMultiple signs in one scene

Inference output across several sign instances in a city street scene.

View original
Street image showing predicted sign regions and confidence annotations.
Model predictionsA second street example

A further inference example, retaining the predicted label and confidence annotations.

View original

Next project / 01

UTSM dynamometer